Skip to content

C++ Diffusion model toolset with Python and Matlab interfaces

License

Notifications You must be signed in to change notification settings

DrugowitschLab/dm

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

52 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

dm

Diffusion model first passage time distribution and sampling library in C++11, with Python and MATLAB interface.

No guarantee is provided for the correctness of the implementation.

The code is licensed under the New BSD License.

Content

The library provides classes and methods, written in C++11, that on one hand compute the first-passage time densities of diffusion models with two absorbing boundaries, and on the other hand draws first-passage time and boundary samples. Each specialised class inherits the abstract DMBase and provides various optimization to compute this density and to sample. The library supports leaky/weighted integration, time-varying drift rates, and time-varying symmetric or asymmetric boundaries.

In addition to the C++11 implementation, a Python and a MATLAB interface are provided. In both cases, the interface chooses between the various different library classes depending on the provided parameters.

The diffusion models assume a drifting and diffusing particle x(t) that starts at x(0) = 0and whose time-course follows

dx = mu(t) dt + sig(t) dW ,

where mu(t) is the current drift, sig(t) is the current diffusion standard deviation, and dW is a Wiener process. Diffusion is terminated as soon as the particle reaches either the upper boundary theta_u(t) or lower boundary theta_t(t). The library requires theta_u(0) > 0 and theta_l(0) < 0. The time at which either boundary is reached is the first-passage time. The associated densities, g_u(t) and g_l(t), are the joint densities over bounds and first-passage times, such that

integral_0^infinity (g_u(t) + g_l(t)) dt = 1 .

In addition to perfect integration, the library provides leaky integration

dx = (- x(t)/tau(t) + mu(t)) dt + sig(t) dW ,

with time constant tau, as well as weighted integration

dx = mu(t) (k mu(t) dt + dW) .

The library provides specialised classes for time-invariant drifts, mu(t) = mu_0 for all t, time-invariant bounds, theta(t) = theta_0 for all t, symmetric bounds, theta_l(t) = - theta_u(t), and leaks and weights (see below). If not otherwise mentioned, a constant unit diffusion variance is assumed, that is sig^2(t) = 1 for all t.

Usage

The library headers and source are in src/. They require a C++ compiler that supports the C++11 (or later) standard.

ExtArray

All diffusion model classes in the library use ExtArray to represent vectors of real numbers. This ExtArray class is used to provide efficient interfacing to external libraries. Its particular feature is that it can be indexed beyond its actual length, in which case it returns a default last value. This is particularly useful for diffusion models that have a drift or bounds that vary until a certain time, and then remain constant. Similarly, time-invariant (i.e. constant) drifts or bounds are just special cases of ExtArray.

Internally, ExtArray encapsultes a C-style array of type double[], wrapped with a C++11 std::shared_ptr. It can operate in two ownership modes:

  • Shared ownership: all copies of a particular ExtArray share ownership of the encapusalted array through std::shared_ptr. As soon as the last instance associated with this array is destroyed, the array itself is deleted. Shared ownership results from calling the constructor with ExtArray(ExtArray::shared_owner(x), ...) where x is of type double[].

  • No ownership: neither copy of a particular ExtArray has ownership of the array. Here, it is important that the array is not deleted before all ExtArray's that are associated with it. The no-ownership state results from calling the constructor with ExtArray(ExtArray::shared_noowner(x), ...) where x is again of type double[]. This type of ownership is particularly useful to avoid copying data when providing interfaces to other languages. Both the MATLAB and the Python interface use this type of ownership extensively.

If a last argument is provided to the constructor of ExtArray, this element is returned for indices beyond the size of the stored array. Otherwise, the array's last element at construction is returned. ExtArray::size() returns the actual size of the stored array, such that x[n] for all n >= x.size() returns the same, last element. This last element is not included in size(), such that an ExtArray of size 0 returns the last element for all indicies n.

DMBase and inherited classes

All diffusion model classes are based on the abstract DMBase class. This class defines the interface to compute first-passage time densities and to draw samples, and provides various factory function to create diffusion models. For time-varying drifts/bounds, the corresponding vectors need to be specified in steps of dt, where dt is provided upon construction.

The currently provided factory functions are:

Create diffusion model with constant or time-varying drift, and constant or time-varying symmetric boundaries

DMBase* dm = DMBase::create(const ExtArray& drift, const ExtArray& bound, value_t dt);

In the above, the drift mu(t) is specified by drift, and the bounds theta_u(t) = -theta_l(t) by bound. dt specifies the time-step of the associated vectors.

Create a diffusion model with constant or time-varying drift, constant or time-varying diffusion variance, and constant or time-varying asymmetric boundaries, with or without leaky integration

DMBase* dm = DMBase::create(const ExtArray& drift, const ExtArray& sig2,
                            const ExtArray& b_lo, const ExtArray& b_up,
                            const ExtArray& b_lo_deriv, const ExtArray& b_up_deriv,
                            double dt[, double invleak]);

In the above, drift is the vector of drifts, sig2 the vector of diffusion standard deviations, b_lo the vector specifying the lower boundary theta_l(t), b_up(t) the vector specifying the upper boundary theta_u(t), b_lo_deriv the vector of time-derivatives of the lower boundary theta_l'(t), and b_up_deriv the vector of time-derivatives of the upper boundary theta_u'(t). dt again specifies the time-step of the associated vectors. The optional invleak specified the inverse leak time constant 1/tau.

For weighted integration, use

DMBase* dm = DMBase::createw(const ExtArray& drift, const ExtArray& bound,
                             double k, double dt);

DMBase specifies the following methods to compute the first-passage time densities:

void pdfseq(int n, ExtArray& g1, ExtArray& g2);
double pdfu(double t);
double pdfl(double t);

pdfseq(n, g1, g2) computes the first-passage time densities g_u(t) and g_l(t) in n steps of dt (provided to the constructor). Both g1 and g2 need to have a capacity of at least n. After calling pwdseq(.), they contain g_u(dt), g_u(2 dt), ..., g_u(n dt) and g_l(dt), g_l(2 dt), ..., g_l(n dt), respectively.

pdfu(t) and pdfl(t) return g_u(t) and g_l(t), respectively. WARNING: by default, they call pdfseq(n,..) with n > t / dt and interpolate between the returned values. If the first-passage time densities need to be computed for both boundaries or multiple times, it is always more efficient to call pdfseq(.) directly. Only the DMConstDriftConstBound and DMConstDriftConstABount classes contain specialised implementations of pdfu(t) and pdfl(t) that are faster than a single call to pdfseq(.).

DMBase provides the following method to draw first-passage time and boundary samples:

DMSample rand(rngeng_t rngeng);

This method returns a single diffusion model sample, where rngeng is and C++11 random number generator engine of type DMBase::rngeng_t. The returned sample is of type DMSample, which provides two methods. The method t() returns the first-passage time of the sample, and the method upper_bound() returns true if the upper boundary was reached, and false otherwise.

In addition to the above, DMBase provides the following convenience method:

void DMBase::mnorm(ExtArray& g1, ExtArray& g2);

This method normalises the mass such that dt (sum(g1) + sum(g2)) = 1, and additionally makes sure that g1(t) >= 0 and g2(t) >= 0 for all t. It does so by adding mass to the last elements of g1 and g2 such that the ratio sum(g1) / (sum(g1) + sum(g2)) remains unchanged.

The factory functions creates an instance of one of the following classes:

DMConstDriftConstBound: constant drift, constant symmetric bounds.

DMConstDriftConstABound: constant drift, constant asymmetric bounds.

DMConstDriftVarBound: constant drift, time-varying symmetric bounds.

DMVarDriftVarBound: time-varying drift, time-varying symmtric bounds.

DMWVarDriftVarBound: time-varying drift, time-varying symmetric bounds, weighted integration.

DMGeneralDeriv: time-varying drift, time-varying variance, time-varying asymmetric bounds.

DMGeneralLeakDeriv: time-varying drift, time-varying variance, time-varying asymmetric bounds, leaky integration.

Interfaces

See matlab/README.md and python/README.md for a description of the corresponding interface.

References

In general, the library computes the first-passage time densities by finding the solution to an integral equation, as described in

Smith PL (2000). Stochastic Dynamic Models of Response Time and Accuracy: A Foundational Primer. Journal of Mathematical Psychology, 44 (3). 408-463.

For constant drift and bounds, it instead uses a much faster method, based on an infinite series expansion of these densities, as described in.

Cox DR and Miller HD (1965). The Theory of Stochastic Processes. John Wiley & Sons, Inc.

and

Navarro DJ and Fuss IG (2009). Fast and accurate calculations for first-passage times in Wiener diffusion models. Journal of Mathematical Psychology, 53, 222-230.

Samples are in the most general case drawn by simulating trajectories by the Euler–Maruyama method. For diffusion models with constant drift and (symmetric or asymmetric) boundaries, the following significantly faster method based on rejection sampling is used:

Drugowitsch J (2016). Fast and accurate Monte Carlo sampling of first-passage times from Wiener diffusion models. Scientific Reports 6, 20490; doi: 10.1038/srep20490.

About

C++ Diffusion model toolset with Python and Matlab interfaces

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Contributors 3

  •  
  •  
  •